This dissertation presents a neural network-based method for estimating the element-level stiffness parameters of a complex structural system, and for detecting the locations and severities for element-level damages and joint damages of structures. A neural network-based substructural identification is presented to estimate the stiffness parameters, such as submatrix scaling factors, particularly in the case of noisy and incomplete measurement of modal data. The concept of substructuring is adopted for local identification. The Latin hypercube sampling and the component-mode-synthesis methods are used for efficient generation of patterns to train and test the neural network. The modal strain energy coefficients are found to be effective indicators in selecting vibration modes for input to the neural networks. The numerical example analyses are carried out on a 2-span truss and a multi -story frame. The noise injection learning using noise at a similar level to the measurement error is found to be effective in order to improve the estimation accuracy.
Secondly, a method was presented for estimating the element-level damages. The damage estimation procedure is divided into the damage location by the damage indicator method and the severity identification by the neural networks in order to overcome the problems related to a large number of probable damaged members. Based on the results of example analyses using simulated data, the proposed combined usage of the damage indicator method and the neural networks is judged to be a very effective and efficient way.
Lastly, a method is proposed to estimate the structural joint damages from the mode shape information using neural networks. The connection stiffness is represented by the stiffness of a zero -length rotational spring at the end of a beam element to model the beam-to- column connection in a steel frame structure. The joint damage severity is defined as the reduction ratio of the connection stiffness. An exa...